Full Strong Branching (FSB)

• Full strong branching is an exhaustive strong branching method that probes every fractional candidate variable by solving both child LP relaxations to compute dual-bound improvements.
• FSB uses scoring rules such as product, linear, and ratio to assess local gains, balancing precise solution progress with the computational burden of multiple LP solves.
• FSB is pivotal in reducing tree sizes in structured problems like vertex cover and has been generalized for nonlinear and spatial branch-and-bound settings with tailored modifications.

Full strong branching (FSB) is the classical exhaustive form of strong branching in branch-and-bound for mixed-integer programming. At a node, every eligible branching variable—typically every variable that is fractional in the current LP relaxation—is probed by solving the LP relaxations of both child nodes induced by the two branchings, and the final variable is chosen from a score computed from the two resulting dual-bound improvements. In this “full” form, strong branching evaluates all candidates rather than a subset, so its decision quality is typically very high, but its per-node cost is correspondingly large because it requires two child LP solves per candidate (Dey et al., 2021, Shah et al., 13 Jul 2025).

1. Definition and canonical scoring

In a minimization MILP, let $N$ be the current node, let $z_N^{LP}$ be the LP relaxation value at $N$, let $PB$ be the incumbent primal bound, and let $C$ be the set of fractional candidates $x_i$ in the current LP solution. FSB probes both child nodes for every $i \in C$ by temporarily fixing $x_i=0$ and $x_i=1$ and solving the two LP relaxations to optimality. If $B_i^0$ and $z_N^{LP}$0 denote the resulting LP relaxation values, the local LP gains are

$z_N^{LP}$1

with infeasible children treated as $z_N^{LP}$2 (Shah et al., 13 Jul 2025).

In a maximization formulation, the same rule is written with reversed signs. If $z_N^{LP}$3 is the LP optimum at node $z_N^{LP}$4 and $z_N^{LP}$5 are the child LP optima after branching on $z_N^{LP}$6, then

$z_N^{LP}$7

FSB chooses a variable maximizing a score computed from the pair of gains (Dey et al., 2021).

The most widely used score is the product rule,

$z_N^{LP}$8

with a small $z_N^{LP}$9 to stabilize ties and zero gains. Alternative classical FSB scores include the linear rule

$N$0

with $N$1 recommended in the literature, and the Le Bodic–Nemhauser ratio rule, defined through the unique root $N$2 of

$N$3

followed by $N$4 (Shah et al., 2024, Dey et al., 2021).

The defining feature of FSB is therefore not a single score formula but exhaustive local evaluation: every candidate variable is tested on both branches before any branching decision is committed.

2. Algorithmic role and computational cost

The operational workflow of FSB is straightforward. At a node with $N$5 fractional candidates, the solver solves $N$6 child LP relaxations, computes a scalar score for each candidate from the two dual gains, selects the highest-scoring variable, and branches. Because the full rule evaluates all candidates, its per-node overhead is much larger than that of pseudo-cost or most-infeasible branching, but this overhead is often offset by much smaller trees (Dey et al., 2021, Mexi et al., 2023).

This computational asymmetry produced a large family of practical surrogates. Reliability branching combines pseudo-cost estimates with limited strong branching, using strong branching only for variables whose pseudo-costs are not yet reliable. Pseudo-costs are updated from unit gains per unit fractional movement, while static lookahead limits and LP-iteration budgets restrict how much strong branching is performed at a node. In one reliability-branching implementation, pseudo-costs become “reliable” after at least $N$7 observations per branch, strong-branching seeding is limited to at most $N$8 variables per node, and each strong-branching LP is limited to $N$9 dual-simplex iterations (Shah et al., 13 Jul 2025). Modern solver implementations also cap strong branching with working limits such as a lookahead parameter and an LP-iteration budget, precisely because unrestricted FSB can dominate runtime (Mexi et al., 2023).

Two notable attempts to manage the cost-quality trade-off depart from pure breadth-first probing. “Narrow gauge” branching evaluates fewer candidates at the current node but uses a small-depth look-ahead tree, together with pre- and post-winnowing, to extrapolate the downstream effects of a branching decision. “Analytical branching” instead tries to replace some child LP solves by analytically chosen pseudo-cost information from structurally nearby nodes in the tree (Glover et al., 2015). A different response is probabilistic lookahead strong branching, which fits a distribution to observed geometric-mean dual gains and stops strong-branching evaluation when the expected reduction in total tree size from one more sample no longer exceeds the sampling cost. In SCIP experiments on MIPLIB 2017, this dynamic stopping rule improved both runtime and node count over a static lookahead baseline, and early stopping was triggered in about $PB$0 of strong-branching calls (Mexi et al., 2023).

A persistent practical interpretation is that FSB is the highest-fidelity local branching oracle, while most modern branching rules are controlled approximations of it.

3. Provable strengths and weaknesses

The most developed theoretical analysis of FSB is for $PB$1–$PB$2 ILPs, especially vertex cover. For vertex cover, the LP relaxation satisfies the Nemhauser–Trotter half-integrality and persistency properties: extreme points are half-integral, and variables that are integral in an LP-optimal solution can be fixed without loss of optimality. The analysis shows that FSB recursively and consistently exploits this persistency throughout the branch-and-bound tree. If $PB$3 denotes the additive integrality gap, then for any vertex cover instance $PB$4,

$PB$5

for the branch-and-bound tree generated by FSB with worst-bound node selection and depth-based tie-breaking (Dey et al., 2021).

The same study also gives a favorable computational benchmark against optimal trees. Using a dynamic program over faces of $PB$6 to compute optimal branch-and-bound tree size under worst-bound node selection, the authors found that across all tested instance families the geometric mean tree size was $PB$7 for the optimal tree, $PB$8 for FSB with the linear score, and $PB$9 for FSB with the product score. Their principal computational conclusion was that, on those tested families, FSB remained within a factor of two of the optimal branch-and-bound tree (Dey et al., 2021).

The positive picture is not universal. The same work identifies an extended-formulation construction based on Bodur–Dash–Günlük in which any tree that branches only on currently fractional variables—including FSB—must have size at least

$C$0

whereas another branch-and-bound tree for the equivalent formulation has size at most

$C$1

This establishes that FSB can be exponentially worse than a different branching strategy when the latter is allowed to exploit auxiliary variables or branch on variables that are not fractional in the current LP solution (Dey et al., 2021).

A common misconception is therefore that FSB is “near-optimal” in an unrestricted sense. The theoretical record is more precise: FSB is exceptionally strong on some structured classes, notably vertex cover, but it is not universally optimal and can be exponentially suboptimal on carefully constructed families.

4. Modified FSB scores and the use of additional information

A recent line of work argues that classical FSB uses only local LP gains and thereby misses two kinds of information: the effect of pruning by the current primal bound, and global asymmetry in how the tree tends to terminate. In the modified formulation, the additive primal–dual gap at node $C$2 is

$C$3

and raw gains are replaced by “efficacious gains”

$C$4

This caps gains that exceed what can actually improve the global dual bound once a child would already be pruned by the incumbent (Shah et al., 13 Jul 2025).

The same work then augments FSB with global last-assignment asymmetry statistics. If $C$5 denotes the set of integral or infeasible leaves discovered so far, and $C$6 count how often the last fixing entering such a leaf was $C$7 or $C$8, the normalized asymmetry is

$C$9

With a scale parameter $x_i$0, this produces exponents $x_i$1 that bias scoring toward the branch statistically less prone to immediate pruning. The resulting modified score is

$x_i$2

where $x_i$3 and $x_i$4 (Shah et al., 13 Jul 2025).

The framework distinguishes several variants. LA-SB uses asymmetry statistics from integral or infeasible leaves. R-LA-SB uses the same last-assignment statistic over all leaves to rebalance trees when pruning is mainly by bounds. PA-LA-SB switches between the two via the observed fraction $x_i$5, using LA-SB when $x_i$6 and R-LA-SB otherwise. The prescribed defaults are $x_i$7, $x_i$8, sample threshold $x_i$9, and asymmetry updates every $i \in C$0 nodes (Shah et al., 13 Jul 2025).

On MIPLIB 2017 Benchmark instances, these modifications reduced mean tree sizes for solved cases by $i \in C$1–$i \in C$2 and lowered the remaining gap for unsolved cases by $i \in C$3–$i \in C$4. The same ideas also transferred to reliability branching, where improved scores reduced mean tree sizes by $i \in C$5–$i \in C$6 on solved instances and lowered the mean gap by $i \in C$7–$i \in C$8 on unsolved instances, depending on primal bound quality (Shah et al., 13 Jul 2025).

These results do not replace FSB’s core mechanism; rather, they reinterpret the score as an estimate of realizable global progress instead of raw local LP improvement alone.

5. Interaction with linear relaxations and cutting planes

A second common misconception is that tightening the LP relaxation must reduce the size of an FSB tree. This is false in general. Monotonicity, in this context, means that for two relaxations $i \in C$9 with the same integer hull and for a fixed branching rule $x_i=0$0, one always has

$x_i=0$1

for all admissible objectives $x_i=0$2. It has been proved that any branching rule that branches only on variables fractional in the current LP solution is non-monotonic. Because FSB is commonly implemented with precisely that candidate restriction, FSB is non-monotonic (Shah et al., 2024).

The same work constructs a family in which adding a single valid inequality improves the LP bound yet causes an exponential increase in FSB tree size. The result holds for the product, linear, and ratio FSB scores. The mechanism is that the tighter relaxation changes which variables are fractional and which local gain pattern FSB sees; the branching sequence then becomes less favorable even though the relaxation itself is stronger (Shah et al., 2024).

Empirical evidence indicates that this is not merely a pathological theorem. In random multidimensional knapsack instances tightened by cover cuts, $x_i=0$3 of individual cuts increased FSB tree size. When all separating cover cuts were added simultaneously, tree size increased in $x_i=0$4 instances; in those cases the total root-gap closure was at most $x_i=0$5. On MIPLIB 2017 instances with SCIP root cuts, many instances showed tree-size increases as cut rounds increased. The broad pattern was that when the gap closed by cuts was small, changes in FSB tree size were difficult to predict and often adverse, whereas sufficiently large gap closure generally led to smaller trees. Cut depth was not a reliable predictor of the effect on FSB tree size (Shah et al., 2024).

This suggests that FSB is highly sensitive to the geometry of the current relaxation, not merely to its nominal bound quality.

6. Generalizations beyond binary MIP

In nonlinear and spatial branch-and-bound, the FSB template survives but no longer suffices by itself. For a continuous variable, branching requires not only choosing the variable but also choosing the branching point. Earlier strong-branching extensions for nonlinear problems predefined one point per candidate variable and then solved $x_i=0$6 relaxations, where $x_i=0$7 is the number of branching candidates. For QCQPs, extreme strong branching (ESB) generalizes the FSB rationale by evaluating multiple branching points per variable via binary search over the current interval, jointly selecting both variable and point from a strong-branching-style score,

$x_i=0$8

and exploiting bound tightening as a byproduct of the search (Dey et al., 23 Oct 2025).

On benchmark QCQPs from MINLPLib, including pooling and water management instances, ESB was the only branching strategy that solved all $x_i=0$9 solved instances under $x_i=1$0 seconds; its average solve times over those instances were $x_i=1$1 and $x_i=1$2, compared with $x_i=1$3 and $x_i=1$4 for the basic predefined-point rule and $x_i=1$5 and $x_i=1$6 for the balance rule. On $x_i=1$7 unsolved instances, the average remaining optimality gap was $x_i=1$8 for ESB, compared with $x_i=1$9 for the basic rule and $B_i^0$0 for the balance rule. The appendix also reports arithmetic mean explored nodes of $B_i^0$1 for ESB versus $B_i^0$2 for the basic rule, even though ESB solved roughly four times as many LP relaxations as the basic rule (Dey et al., 23 Oct 2025).

The finite-element-model updating study in the same paper reports a similarly strong contrast with commercial software: Gurobi solved $B_i^0$3 instances with $B_i^0$4 and $B_i^0$5, whereas the ESB branch-and-bound solved $B_i^0$6 with $B_i^0$7 and $B_i^0$8. Across all instances, arithmetic mean node counts were $B_i^0$9 for Gurobi and $z_N^{LP}$00 for ESB (Dey et al., 23 Oct 2025). A plausible implication is that, in nonlinear settings, the core FSB idea—look ahead on both children and score the pair jointly—remains effective, but only after the branching-point decision is made part of the branching rule itself.

Learning-based studies in spatial branching provide a further qualification. In continuous polynomial optimization with RLT relaxations, a perfect strong-branching-imitation oracle for variable choice (“BVar”) was found to perform worse overall than both static instance-level rule selection (“ORule”) and node-level rule selection (“BRule”) on time, gap, and the pace of lower-bound improvement, even though BVar often produced smaller geometric mean node counts. On the reduced $z_N^{LP}$01-instance set, representative geometric means were: ORule, Pace $z_N^{LP}$02 and Nodes $z_N^{LP}$03; BRule, Pace $z_N^{LP}$04 and Nodes $z_N^{LP}$05; ML-Q-RF, Pace $z_N^{LP}$06 and Nodes $z_N^{LP}$07; BVar with tie threshold $z_N^{LP}$08, Pace $z_N^{LP}$09 and Nodes $z_N^{LP}$10 (González-Rodríguez et al., 2024). This suggests that, in nonlinear spatial branching, immediate dual-bound improvement can be too myopic to serve as a universally dominant branching signal.

Across these extensions, FSB retains its status as the reference exhaustive look-ahead rule, but its generalization requires either richer scoring, richer candidate parametrization, or explicit control of the computational burden that exhaustive probing creates.